Received May 10, 2020, accepted May 13, 2020, date of publication May 18, 2020, date of current

version June 1, 2020.

Digital Object Identifier 10.1109/ACCESS.2020.2995298

Grid-Forming Power Converters Tuned Through

Artificial Intelligence to Damp Subsynchronous
Interactions in Electrical Grids
GREGORY N. BALTAS 1 , (Student Member, IEEE), NGOC BAO LAI 1,2 , (Student Member, IEEE),
LEONARDO MARIN2 , (Student Member, IEEE), ANDRÉS TARRASÓ2 , (Student Member, IEEE),
AND PEDRO RODRIGUEZ1,2 , (Fellow, IEEE)
1 Loyola Institute of Science and Technology (Loyola.TECH), Universidad Loyola Andalucía, 41704 Seville, Spain
2 Renewable Electrical Energy Systems (SEER), Technical University of Catalonia, 08222 Terrassa, Spain
Corresponding authors: Gregory N. Baltas (ngbaltas@uloyola.es) and Ngoc Bao Lai (nblai@uloyola.es)
This work was supported in part by the European Commission under Project FLEXITRANSTORE—-H2020-LCE-2016-2017-SGS-
774407, and in part by the Spanish Ministry of Science under Project ENE2017-88889-C2-1-R.

ABSTRACT The integration of non-synchronous generation units and energy storage through power
electronics is introducing new challenges in power system dynamics. Specifically, the rotor angle stability
has been identified as one of the major obstacle with regards to power electronics dominated power systems.
To date, conventional power system stabilizer (PSS) devices are used for damping electromechanical
oscillations, which are only tuned sporadically leading to significant deterioration in performance against
the ever-changing operating conditions. In this paper, an intelligent power oscillation damper (iPOD) is
proposed for grid-forming converters to attenuate electromechanical inter-area power oscillation. In partic-
ular, the iPOD is applied to a synchronous power controller (SPC) based grid-forming power converter to
increases gain of the active power control loop at the oscillatory frequency. Predictions regarding the mode
frequency, corresponding to the current operating points, are given by an artificial intelligence ensemble
model called Random Forests. The performance of the proposed controller is verified using the two area
system considering symmetrical fault for random operating points. In addition, a comparison with PSS
installed in each generator reveals the individual contribution with respect to the inter-area mode damping.

INDEX TERMS Artificial intelligence, ensemble modelling, inter-area power oscillation, random forests,
synchronous power controller.

I. INTRODUCTION
Renewable-based Energy Systems (RES) and Energy Storage
Systems (ESS) are widely adopted to reduce green house
gas emissions and improve efficiency. These systems, due to
their incompatibility with synchronous frequency (50/60 Hz),
require power electronic devices to act as an intermedi-
ate before connection with the rest of the network is pos-
sible. Consequently, as the presence of power electronics
increases, the power system dynamics are affected intro-
ducing new challenges regarding power system stability [1].
In fact, ENTSO-E has identified the angular stability of
power systems as one of the key challenges related to High
Penetration of Power Electronic Interfaced Power Sources
(HPoPEIPS) [2].
The associate editor coordinating the review of this manuscript and
approving it for publication was Xiaorong Xie . parameters that need to be tuned for attenuating system
Conventionally, the PSS has been used to damp low fre-
quency oscillations in a system by regulating the synchronous
generator’s excitation voltage.Among the standardized PSS
schemes, the classical PSS1 controller consists of three main
blocks: gain, washout filter and phase compensation. Consid-
ering inter-area modes, the time constant of the washout filter
needs to be carefully chosen, since it captures the low fre-
quency oscillations [3]. The PSS1 uses a single input signal,
typically the speed deviation 1ω. More sophisticated designs
are the PSS2B and PSS4B controllers. In PSS2B, speed
transducers are added to improve reactive power modulation
during mechanical power changes and eliminate torsional
forces that can harm the generator shaft. PSS4B builds on
top of the PSS2B by incorporating low, intermediate and high
frequency bands to broaden the range of targeted modes [4].
One major limitation of the PSS is the high number of

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 8, 2020 93369
 G. N. Baltas et al.: Grid-Forming Power Converters Tuned Through Artificial Intelligence
oscillations [5]. For instance, a typical PSS has 8 parameters:
gain, time constant of wash out filter, and four constants for
lead-lag compensator. These parameters usually are tuned at
commissioning of the power plant and remain fixed despite
the changes of operating conditions [6]. There is an open
discussion in the literature on possible methods for adaptive
PSS and optimal tuning [3]. This becomes even more relevant
now, as power electronics can alter the oscillatory profile of
a given system, changing or creating new modes.
Currently, the number of synchronous generators is slowly
phased out [7] and as a result so will the number of PSS. Yet,
power electronic interfaced generation units can provide a
fast response to the grid and assist the system in maintaining
its stability in case of a disturbance [8]. Such grid friendly
response can be produced by means of the currently well-
known grid-forming power converters. In general, the control
system of a grid-forming converter can either be based on
droop [9] or on virtual synchronous machine (VSM) [10].
In the former case, the contribution of the power converter is
usually limited to primary frequency and voltage regulation.
To add grid-supporting functionalities, e.g. inertia emulation
and power oscillation damping, the VSM implementation can
be adopted in the grid-forming power converter [11].
Among the various VSM strategies, reported in the lit-
erature, the Synchronous Power Controller (SPC) has been
widely used mainly due its simple yet effective control
structure [12]. By employing the SPC, power converters can
provide virtual damping and synthetic inertia to the grid,
increasing its dynamic flexibility to improve local and inter-
area stability [13], [14]. In fact, the SPC with multiple power
loop controllers (PLC) has been presented in [15]. By com-
bining a band-pass filter with a modified swing equation,
the inter-area oscillations at a predefined frequency can be
attenuated. However, to ensure the system in [15] works
effectively, the center frequency of its band-pass filter has
to be tuned properly to match the frequency of the inter-area
oscillation mode, which is usually not accessible due to the
dynamic behavior of the power system.
Recently, data-driven methods based on artificial intelli-
gence (AI) are studied for accurately and efficiently deter-
mining the state of the power system in real time and thereby
avoid instability [16]. Specifically, the relationship between
system variables and stability indexes can be modeled to pre-
dict the post-fault state of the system. For instance, an ensem-
ble approach based on extreme learning machines is adopted
in [17] to predict stability after a disturbance, while a recur-
rent neural network with long short term memory (LSTM)
cells is trained in [18] for the same task considering time
dependencies in the data. Furthermore, prediction of damping
state (i.e. well or poorly damped) is proposed by [19] using
decision trees and by [20] using neural networks.
Driven from the above, this paper presents an intelligent
Power Oscillation Damper (iPOD) that enables the SPC to
adapt for maximum -oscillatory- mode attenuation. The iPOD
is based on an ensemble AI predictor called Random Forests
(RF), which is trained to predict the real and imaginary
93370
components of the targeted mode of oscillation as it moves
according to the current operating conditions. The main con-
tribution of the system, presented in the following, is a novel
control scheme that incorporates intelligence by exploiting
the ability of AI to predict the characteristic values of an
oscillatory mode in real time and uses this information to
adaptively tune a dedicated loop of the SPC to damp oscilla-
tions and enhance the system stability i.e. iPOD. In this work,
special focus is put on describing all the steps for developing
the AI models in detail. The proposed approach is verified
under a symmetrical fault for random operating points as
well as against conventional PSS devices using the well-
known two area system described in [21]. The AI predictor
is trained for tracking the inter-area mode between the two
areas, however the same principles can be applied to develop
multiple AI predictors to predict more than one inter-area
modes.
The remainder of this paper is structured as follows.
Section II provides a detail description regarding the test case.
Section III presents the proposed AI-based adaptive SPC
model. Section IV analyse the results and Section V discusses
the conclusions.
II. TEST CASE
A popular power system to analyse the behavior of an inter-
area mode is the two-area system [21], depicted in Fig. 1.
The system consists of two areas with one load (L07, L09),
one shunt capacitor (C7, C9) and two generators (G1-G2,
G3-G4), respectively. Each generator is equipped with IEEE
Type AC4A Excitation System (AVR), steam turbine gov-
ernors (GOV) and PSS type 1. These areas are connected
through a weak tie-line where 400MW are flowing from Area
1 to Area 2. Modal analysis reveals two local modes with
frequencies of 1.05Hz and 1.08Hz in Area 1 and Area2,
respectively, as well as an inter-area mode with frequency of
approximately 0.55Hz.
The specific design of this system not only allows for
studying the behaviour of electromechanical oscillations
between two areas, but also on demonstrating the effective-
ness of control methods (such as the iPOD) for attenuating
them. To demonstrate the impact of the iPOD, a virtual-
synchronous generation unit equiped with an SPC is con-
nected at bus 7, through two step-up transformers. The rated
apparent power of the aforementioned unit is 100MVA, equal
FIGURE 1. Modified two area system.
VOLUME 8, 2020
G. N. Baltas et al.: Grid-Forming Power Converters Tuned Through Artificial Intelligence

FIGURE 2. Block diagram of RF-SPC.

FIGURE 3. Intelligent power oscillation damper (iPOD).

to 3% of the total apparent power of the four synchronous
generators together.

III. iPOD FOR GRID-FORMING CONVERTERS

A. CONTROL STRUCTURE
For mitigating inter-area oscillations, a conventional SPC is
equipped with an iPOD as shown in Fig. 2. Generally, the SPC
consists of four main control blocks, namely, the current
controller, the virtual admittance, the reactive controller, and
the PLC. In the SPC, the reference for the current control loop
is generated by means of a virtual admittance whose transfer
function is
1
Y (s) =
Ls + R
where R and L are the virtual resistance and the virtual
inductance, respectively. As shown in Fig. 2, the virtual elec-
tromotive force eemf is generated by combining the outputs
of the active and reactive control loops through a voltage-
controlled oscillator (VCO) as
eemf = Vref sin(θ).
In (2), the magnitude of the electromotive force is regulated
through the reactive power controller as follows:
Z to be a known parameter. Indeed, the oscillation frequency is
Vref = VBase + kp (Qref − Qm ) + ki (Qref − Qm )dt (3)
where kp and ki are the proportional and integral gains of the
reactive power controller, and Vbased is the nominal voltage.
Likewise, the phase angle θ used in (2) is produced by the
active power loop controller (PLC). In this implementation,
although any other controller structure could be used [8],
the PLC is implemented through the conventional swing
equation whose transfer function is
1 Gp (s) = GPLC (s) + GBPF (s) [k1 GPLC (s) + k2 ]
GPLC (s) =
2Hs + D
where H and D are the inertia constant and the damping coef-
ficient respectively, while ωbase is the grid nominal frequency,
which acts as a feed-forward term to improve the dynamic
performance of the SPC.
Aiming to increase damping to inter-area oscillation
modes, the PLC is incorporated with the proposed iPOD as
(1)
FIGURE 4. Frequency response of the conventional PLC and the
proposed PLC.
shown in Fig. 3. Essentially, the iPOD includes two propor-
tional gains k1 and k2 , and a second-order band-pass filter.
The gain k1 is set to −1 for removing the targeted frequency
(2) component from the signal before acted on the PLC, whereas
k2 defines the damping factor provided by the iPOD at the
specified frequency. The reason behind the choice of a band-
pass filter is that the frequency of the oscillation is considered
predicted by the AI predictor which will be explained in the
following subsections. The transfer function of the band-pass
filter can be given as
Bs
GBPF (s) = (5)
s2 + Bs + ωc2
where ωc is centre frequency and B is passing band. From (4)
and (5), the augmented transfer function of the active power
controller can be derived as follows:
(4)
= GPLC (s) + GBPF (s)Gll (s)
(2BHk2 + 1)s2 + (B + Bk1 + BD k2 )s + ωc2
=
2Hs3 + (D + 2BH )s2 + (2H ωc2 + BD)s + Dωc2
(6)
From (6), the combined frequency response of the PLC
and the iPOD is obtained, as depicted in Fig. 4 for different

VOLUME 8, 2020 93371

 G. N. Baltas et al.: Grid-Forming Power Converters Tuned Through Artificial Intelligence
values of gain k2 . As shown, the iPOD adds a significant
amplification to the active power control loop at the tuned
frequency. Such a high gain implies that the proposed PLC
with iPOD control structure can provide higher damping to
low-frequency inter-area oscillation.
It is worth noting that the augmented transfer function
(6) has two parts. While the first term is conventional PLC,
the second term reflects added effects of the iPOD. In fact,
the second term can be rewritten in a form of a lead-lag
compensator as follows:
2k2 Hs + (k1 + k2 D)
Gll (s) = . (7)
2Hs + D
Equation (7) reveals that the two gains k1 and k2 make it
possible to adjust not only gain but also phase lag of the active
power control loop at the specified frequency. In practice,
the inter-area modes depends on the operating condition of
the electrical grid. Therefore, in order to predict the oscilla-
tion modes in real-time, the AI-based algorithm called Ran-
dom Forest is employed.
B. ENSEMBLE AI PREDICTOR
The structure of the proposed iPOD requires up to date
information regarding the targeted mode’s frequency to pro-
vide additional damping. Therefore, its imperative to develop
a model that will be able to relate the systems operating
condition to the specific movement of the targeted mode.
Although this can be a challenging task, following a purely
mathematical approach, data-driven methods based on AI and
machine learning are able to approximate this relationship
without any knowledge about the inner mechanics of the
system.
Among the plethora of machine learning models, such
as the neural network in [22] to improve damping control
for Static Synchronous Series Compensator, the decision
tree (DT) stands out due to its robustness against feature
scales or types, meaning that it is possible to process pat-
terns with metric and non-metric data at the same time [23].
In addition, the decision making process of the DT is highly
interpretable, e.g. as in Fig. 5, and is able to remove redun-
dant system variables, thus performing an intrinsic feature
selection.
The basic principle of DTs is to classify patterns based on
a series of ‘‘questions’’, where the given answer determines
the next ‘‘question’’ being asked and so on. This top-down
process can be displayed in the form of a tree, consisting
of nodes and branches as illustrated in Fig. 5. The process
begins at the root node and it ends at the leaf nodes, where
the process stops. The ‘‘question’’, more formally known as
the split criterion of the node, resembles an ‘‘if-then-else’’
statement that describes a specific rule. For instance, the root
node of the DT depicted in Fig. 5, reveals a threshold value
for the Voltage magnitude (Vm ) at bus 09. Depending on the
answer the subsequent question can be either about the Power
Factor (PF) at bus 10 or at bus 9, as in (8). This rule based
approach is the core of the high level of transparency, which
93372
FIGURE 5. Feature space segmentation of a regression tree.
is not found in other machine learning models.
(
B09 Yes ⇒ PF B09 ≤ −0.991
Vm ≤ 0.963 ⇒ (8)
No ⇒ PF B10 ≤ 0.987
Generally, the purpose of the splitting criterion is to divide
the feature space into as pure as possible sub-spaces. Hence,
for every candidate split criterion the resulted impurity is
calculated as the mean squared error, see (9), between the
true value y(i) and predicted value ŷt for patterns in subset St
averaged over all training patterns Nt at node t. The predicted
value at node t is calculated according to (10).
1 X (i)
I (t) = (y − yˆt )2 , (9)
Nt
i∈St
1 X (i)
ŷt = (y ) (10)
Nt
i∈St
The efficiency of a split criterion is further validated by cal-
culating the drop of impurity at the ancestor nodes, as in (11).
Here, tY and tN are the two ancestor nodes corresponding to
the Yes and No branches, while NtY /Nt and NtN /Nt are the
portion of samples at each of the these two nodes. Ultimately,
at each node, the final splitting rule is chosen to be the one
that reduces the impurity at the ancestor nodes the most [24].
NtY NtN
1I (t) = I (t) − I (tY ) − I (tN ), (11)
Nt Nt
Despite their straightforward interpretation and intuitive-
ness, DTs tend to overfit the data demonstrating poor gener-
alization. Moreover, errors propagate from higher level nodes
towards lower level nodes [24] resulting in a different DT if
changes to the training data occur.
To overcome the avove limitations, an ensemble AI model
called Random Forest (RF) is adopted, which integrates mul-
tiple DTs into a single predictor, as in Fig. 6. The individual
DTs are trained, based on the principle describe above, for
the same task but separately from each other, using random
subsets of patterns (i.e. system snapshots) and/or features (i.e.
system variables). The random sampling is widely known as
VOLUME 8, 2020
G. N. Baltas et al.: Grid-Forming Power Converters Tuned Through Artificial Intelligence
FIGURE 6. Bagging ensemble - random forests.
bootstrap aggregation (or bagging for short [23]) when the
subsets are not unique. The ensemble paradigm demonstrates
higher performance both in terms of accuracy and general-
ization: a direct consequence of the diversity promoted by
the randomness of the different subsets that are used for
developing each of the individual prediction models.
Frequently, the final output of the ensemble is given
through a majority voting scheme [25] or by averaging
the probabilistic prediction (i.e. soft voting). Nevertheless,
the core of RF are the individual DTs where the development
process is the same, albeit the variations of the feature-pattern
space.
C. ENSEMBLE AI PREDICTOR TRAINING AND TUNING
1) DATABASE DESCRIPTION
To generate a sufficient amount of measurements, an auto-
mated simulation process is designed to parse through a
wide range of probable operating scenarios, and to store the
variables of interest. In other words, this automated process
will formulate a database D ∈ RM ×N , where M ∈ Z+ is the
total number of patterns and N ∈ Z+ is the total number of
features.
DigSilent’s PowerFactory is the selected platform to run
the simulations and Python 3.7 for executing Algorithm 1.
The loading of the system (line 3 of Algorithm 1) is chosen
from (12), where r1,2 ∼ U (0, 1) are random coefficients
drawn from a uniform distribution between the open set 0 and
1, µ is the distribution mean and σ is the standard deviation.
The processing time on a machine with 20 processors and
a single Powerfactory licence requires around 50 minutes
per 1000 simulations. Our final dataset consists of about
22000 patterns and 122 features, i.e. D ∈ R22000×122 . A sum-
mary of the system variables constituting our feature vector
can be found in Table 1.
β = µ + σ −2 log(r1 ) cos (2π r2 )
p
(12)
For each pattern in the dataset there exists a set of eigen-
values that corresponds to a random operating point. From
VOLUME 8, 2020
Algorithm 1 Generate Data-Set
1: input M
2: while m < M do
3: Randomly Set P and Q
4: Calculate Power Flow
5: if Success then
6: Perform Modal Analysis
7: Store Results
8: m←m+1
9: end if
10: end while
TABLE 1. Monitored system variables.
the set of eigenvalues the most dominant mode is extracted,
in this case the inter-are mode. Hence, each pattern is labeled
with the target variable comprised of the real and imaginary
component i.e. y ∈ R22000×2 .
The behaviour of dominant inter-area mode for the random
operating points is revealed in Fig. 7. From the figure it
becomes apparent that there are many instances were the
dominant inter-area mode moved to the, unstable, right hand
side of the real axis. In addition, the frequency of the dom-
inant inter-area mode varies significantly, as the conditions
change, from 0.45Hz up to 0.76Hz. Interestingly, from Fig. 8
it appears to exist a relationship between voltage magnitude
and voltage angle that can be used by the RF to map system
variables to the inter-area mode. For example, almost every
voltage magnitude that corresponds to the unstable inter-area
mode is less than 1 p.u. Similar patterns can be seen in other
locations of the system not presented here for brevity.
2) PREPROCESSING
The simulated environment assumes full observability of
the system, that is, measurements of system variables are
recorded in every location of the system. However, in reality
this can be costly and impractical. In addition correlated
features can impair the ability of a prediction model to make
accurate predictions. Therefore, several methods are used for
reducing the size of the feature vector to include only relevant
and uncorrelated features. In general, the objective of data
preprocessing is mainly two-fold; to improve the accuracy of
the model and its training speed [26].
93373
 G. N. Baltas et al.: Grid-Forming Power Converters Tuned Through Artificial Intelligence
FIGURE 7. Modal analysis results of algorithm 1.
FIGURE 8. Kernel density estimate of voltage magnitude in bus 08 for
random operating conditions.
The inherent ability of DTs and RFs to perform feature
elimination is exploited to analyse the individual contribution
of each feature in the prediction process. Typically, greedy
search algorithms like the Sequential Feature Selection (SFS)
are preferred as they require less time and computational
resources to eliminate features [26]. Fortunately, the size
of the feature vector in hand allows the implementation of
exhaustive search algorithms such as the Recursive Feature
Elimination (REF), which guarantee optimal feature elimi-
nation. In this study, the REF algorithm is used alongside the
RFs to decrease the size of the feautere vector even further to
only 22.
PM
(xi,j − xj )(xi,p − xp )
r = qP i=1 (13)
M 2 (x − x )2
i=1 (xi,j − xj ) i,p p
The RFs are trained in python 3.7 using the scikit-learn
package [27]. Although it is possible to predict both real
and imaginary parts of eigenvalues with a single RF, here
it is preferred to train a separate RF for each component.
Therefore, if performance drops for one of the two target
variables, the problematic model can be still retrained without
losing the predictions for the other variable.
The optimal parameters, valid for both RF models, are esti-
mated using grid search and summarized in Table 2. Notice
that three out of five parameters refer to the development
of the individual DT. Specifically, the minimum patterns per
split define a lower bound to the number of patterns needed
to make a split, the maximum features define the number
of features to be considered for the split criterion at each
node, while the maximum depth sets an upper bound to the
DT growth. Finally, the Bootstrap implies feature and pat-
tern sampling selection is done with replacement. The RF is
evaluated using a 10-fold cross validation, recording both the
R2 score, in (14) and Mean Absolute Error (MAE), in (15).
The former measures how well the features explain the target
variable with values ranging from 0 (worse) to 1 (best).
Pn
(yi − ŷi )2
R2 (y, ŷ) = 1 − Pi=1
n 2
(14)
i=1 (yi − ȳ)
npatterns −1
1 X
yi − ŷi  .
 
MAE(y, ŷ) = (15)
npatterns
i=0

FIGURE 9. Data preprocessing workflow.

The typical workflow followed for processing a dataset
consists of three steps: filter, select and re-scale as in Fig. 9.
However, due to the robustness of DTs to features of different
scale and type last step can be omitted, as indicated by the
3) PERFORMANCE ANALYSIS
Learning curves are showing the performance of a machine
learning model as a function of the training set’s size.
TABLE 2. Random forests parameters.

dashed lines in Fig. 9.

The highly correlated features are filtered out by using
a minimum acceptance threshold and Pearson’s correlation
coefficient, as in (13), where {j, p ∈ [1, . . . , N ] | j 6 = p}
are feature indexes while the bar above the variable denotes
the sample mean. The uncorrelated features comprising
the feature vector reduced from 122 to just 36 features.

93374 VOLUME 8, 2020

G. N. Baltas et al.: Grid-Forming Power Converters Tuned Through Artificial Intelligence

Particularly, Fig. 10 depicts the learning curve of the RF

model developed to predict the ={λ} component of inter-
area mode. The figure clearly shows that after 8000 training
patterns there is no significant gain in the performance of
the RF model. Hence, by using fewer training patterns we
can limit the time spent for fitting the model to just 30s.
Furthermore, the cross-validation score seems to converge
to a certain value indicating that further addition of training
data cannot improve the model. Similar behaviour can be
viewed for the corresponding model of the <{λ} component
not shown here due to space limitations.
In addition to the metrics described in the previous section,
both training and testing CPU processing times are consid-
ered. The cross validated accuracy and speed of <-RF and
=-RF are listed in Table 3. The tables prove the high accuracy
of RF to predict both the real and imaginary values of the
inter-area mode. The error of the <-RF is slight lower than
the one calculated for =-RF however the difference is quite
small. On average a single RF model needs approximately
50s for training and 0.7s on a high performance computer
using 32 CPUs.

TABLE 3. Random forests performance.

FIGURE 10. Learning curves on ={λ} component.

Overall, both RFs demonstrate high accuracy on predicting

the inter-area mode using only 22 system variables. Con-
cretely, RFs can provide accurate information about the inter-
area mode such as damping, frequency and damping ratio.
Although the iPOD is designed for tracking and attenuating
the inter-area mode between the two areas, the same princi-
ples can be applied to develop the iPOD for more than one
inter-area modes. This implies that the iPOD will be modified
by integrating multiple AI predictors (or a single multi-output
AI predictor) and multiple band pass filters [15] as dictated by
the number of inter-area modes in the system. Multiple inter-
area mode attenuation is however, out of the scope of this
paper. The next step is to verify that this information allows
the iPOD to realize mode attenuation and enhance system
stability.
IV. VERIFICATION
The intelligence of the iPOD scheme, proposed in this paper,
for the SPC is based on the premises that a machine learning
model, like RF, can predict in real time the dominant mode
real and imaginary components. Specifically, the SPC can
utilize the prediction about the mode’s frequency to re-tune
the iPOD band-pass filter to adapt in the new operating
conditions. The gains of the iPOD are chosen as k1 = −1
and k2 = 0.018. This section, aims to verify that the iPOD
can improve the attenuation of power oscillations in the
system under contingencies given a random operating point
(see Section III-C1), and compare it to the conventional PSS
performance in the well-known two area system used as a
reference [21].
A. SYSTEM RESPONSE TO DISTURBANCE
To excite the inter-area mode a 3-phase fault is defined in the
middle of the upper transmission line that connects bus 7 and
1 . The fault is cleared by opening the breakers at
8 i.e. L7−8
both ends and re-closing them after 100ms. As mentioned,
loads in areas 1 and 2 are set randomly to move the system
at a new operating point. The snapshot of the system at time
tm (before fault occurrence) is fed into RF, which makes a
prediction about the inter-area to tune the SPC.
Note that in this study communication delays have not
been considered, hence all measurements at tm are in sync.
In reality, however, this might not be the case. Typically,
a data storing and preprocessing module handles the infor-
mation flow before they are passed on for further processing
[28].Communication protocols, such as the well known IEC
61850, which support timestamping are used for ensuring that
measurements can be sorted according to their corresponding
time tags [29], [30].
Illustrated in Fig. 12 is the frequency recorded at buses 6
and 10 of areas 1 and 2, respectively. As a comparison the

VOLUME 8, 2020 93375

 G. N. Baltas et al.: Grid-Forming Power Converters Tuned Through Artificial Intelligence

FIGURE 11. Frequency response in area 1 for a random operating point

and a 3-phase fault in L17−8 .

FIGURE 13. Modal analysis for random operating points.

FIGURE 12. Frequency response in area 2 for a random operating point

and a 3-phase fault in L17−8 .

system response with the proposed SPC equipped with the

iPOD is plotted against the base case (i.e. without SPC) and
the SPC-only case (i.e. without online tuning and attenuation
of low-frequency power oscillations).
The inherent capability of the regular SPC to damp oscil-
lations is revealed in the frequency response of both areas
[31]. Yet, by predicting the mode frequency and re-tune the FIGURE 14. Instantaneous active power of generators in area 1 for a
random operating point and a 3phase fault in L17−8 .
iPOD parameters the damping can be enhanced even further
when the SPC+iPOD is used, as it is shown in Fig. 12, forcing
oscillations to decay faster. TABLE 4. On-fault instantaneous peak power output (in MW).
Concretely, the stochastic variations of loads cause the
inter-area mode to vary in both frequency and damping,
therefore, it is vital to be able to monitor these changes.
The SPC can significantly improve damping of inter-area
modes as shown in Fig. 13 depicting the results of modal
analysis for random operating points. It is worth noting that
the effectiveness of the iPOD scheme reflects the accuracy of
the RF predictions.
As expected, the generators in area 2 are oscillating with
B. OSCILLATION DAMPING IMPACT OF iPOD TO higher amplitudes than those in area 1. Inter-area modes can
SYNCHRONOUS GENERATORS result in power fluctuations, which can stress the structural
In the two area system, the load L09 in area 2 is approxi- integrity of generator units. By integrating the iPOD in the
mately 1.5 times larger than L07 of area 1. As a consequence, conventional SPC, the system power oscillations decay at
area 2 relies on imported power to maintain power balance. a higher rate reducing the stress on generator shafts. This
The power dependence of area 2 can be seen by analysing becomes even more obvious in area 2 where a 100MW oscil-
the power oscillations in each generator. Specifically, these lation can damage generators 3 and 4. Furthermore, the output
responses during the symmetrical fault in L7−8 1 for a random power excursions during the on-fault period are reduced for
operating point are depicted in Fig. 14 for area 1 and in Fig. 15 all generators but generator 2, where there exists a slight
for area 2. increase of peak power as listed in Table 4. Nevertheless, it is

93376 VOLUME 8, 2020

G. N. Baltas et al.: Grid-Forming Power Converters Tuned Through Artificial Intelligence
FIGURE 15. Instantaneous active power of generators in area 2 for a
random operating point and a 3phase fault in L17−8 .
TABLE 5. Parameters of PSS.
TABLE 6. Averaged modal analysis results per control scenario.
worth emphasizing that the improved damping from the SPC
based on iPOD is achieved with only 1/7 of the synchronous
generator apparent power.
C. iPOD COMPARISON WITH PSS
Commonly, the PSS is used to provide damping of critical
modes in a system. This section compares the individual
contribution of each PSS device with the corresponding con-
tribution provided by the SPC augmented by the iPOD. For
the purpose of this study, the PSS parameters are considered
fixed and their values are set following the methodology
described in [32] and listed in Table 5. Concretely, random
points are generated following the aforementioned process
and for each control scenario modal analysis of the system
is performed. The results, averaged over all random points,
are listed in Table 6.
Clearly, PSS1 and PSS3, on average, possess the highest
contribution towards inter-area power oscillation damping.
However, damping provided by PSS1 and PSS3 varies signif-
icantly between each case. Particularly, PSS1 fails to provide
damping for the inter-area mode for a random operating as
VOLUME 8, 2020
FIGURE 16. Eigenvalues of system per control scenario.
revealed in Fig. 16. As opposed, the SPC augmented by
iPOD seems to provide less damping than the previous two
PSS devices. However, as Fig. 16 shows, this amount does
not fluctuate much highlighting the damping capabilities of
the SPC with iPOD. Furthermore, it also dominates over
the contribution of both PSS2 and PSS4, which due to poor
tuning are unable to damp the inter-area mode corresponding
to the random operating points. Note that the nominal power
of the SPC is considerably lower to the nominal power of the
synchronous generators.
V. CONCLUSION
The connection of energy systems through power electronic
devices is adding to the complexity of power systems by
altering the overall behaviour of power systems. Rotor angle
stability is of primary concern as the presence of power elec-
tronics in power systems is rising. Conventional PSS are per-
forming acceptably for attenuating oscillatory modes, how-
ever, only for a narrow range of operating conditions. Power
converters controlled through grid-forming techniques, e.g.
a SPC augmented by an iPOD as in this paper, can provide
an additional source for damping power oscillations with the
use of data and AI.
Random Forest is a powerful machine learning model,
which exploit the inherent diversity of combing individual
decisions trees trained with random subsets of the original
dataset. In this paper, two random forests have been devel-
oped for predicting the real and imaginary component of the
inter-area mode in the two are system. These predictions are
used to tune the iPOD added to SPC for improving damping
of the targeted inter-area mode. As it has been proven from
thorough analyses, the SPC with iPOD can enhance the stabil-
ity of a power system by increasing the damping of the inter-
area mode. In the test presented in this paper, the response
of recorded frequency demonstrated significant improvement
when SPC is augmented by the iPOD scheme. In contrast to
conventional PSS, the iPOD does not have a phase lag and
93377
 G. N. Baltas et al.: Grid-Forming Power Converters Tuned Through Artificial Intelligence
only the frequency of the oscillation is required. Therefore,
reducing the number of parameters needed to be tuned for
attenuating the inter-area mode.
Furthermore, the proposed scheme is shown to be robust
against network variations, performing equally and even
better than the conventional PSS in the system. However,
the proposed control scheme should not be considered as
an adversary of PSS but rather as a complementary ally
for improving the performance of the power systems of the
future, which will be dominated by power electronics. The
increased utilization of grid-forming converters for additional
functionalities other than their primary purpose will increase
flexibility and robustness in the ever-changing power system
environment without the necessity for additional investments.
ACKNOWLEDGMENTS
This work was supported in part by the European Com-
mission under Project FLEXITRANSTORE—H2020-LCE-
2016-2017-SGS-774407, and in part by the Spanish
Ministry of Science under Project ENE2017-88889-C2-1-R.
Any opinions, findings and conclusions or recommendations
expressed in this material are those of the authors and do not
necessarily reflect those of the host institutions or funders.
REFERENCES
[1] L. Meegahapola and D. Flynn, ‘‘Impact on transient and frequency stability
for a power system at very high wind penetration,’’ in Proc. IEEE PES Gen.
Meeting, Jul. 2010, pp. 1–8.
[2] High Penetration of Power Electronic Interfaced Power Sources
(HPoPEIPS) ENTSO-E Guidance Document for National Implementation
for Network Codes on Grid Connection, ENTSO-E, Brussels, Belgium,
2017.
[3] Z. A. Obaid, L. M. Cipcigan, and M. T. Muhssin, ‘‘Power system
oscillations and control: Classifications and PSSs’ design methods:
A review,’’ Renew. Sustain. Energy Rev., vol. 79, pp. 839–849, Nov. 2017.
[Online]. Available: http://www.sciencedirect.com/science/article/pii/
S1364032117307499
[4] I. Kamwa, R. Grondin, and G. Trudel, ‘‘IEEE PSS2B versus PSS4B:
The limits of performance of modern power system stabilizers,’’ IEEE
Trans. Power Syst., vol. 20, no. 2, pp. 903–915, May 2005.
[5] IEEE Recommended Practice for Excitation System Models for Power
System Stability Studies, IEEE Standard 421.5-2016 (Revision IEEE Std
421.5-2005), Aug. 2016, pp. 1–207.
[6] F. Rashidi and M. Rashidi, ‘‘Robust and adaptive tuning of power
system stabilizers using artificial neural networks,’’ in Innovations in
Applied Artificial Intelligence. Berlin, Germany: Springer, Jul. 2004,
pp. 1023–1032.
[7] J. Fang, H. Li, Y. Tang, and F. Blaabjerg, ‘‘On the inertia of future more-
electronics power systems,’’ IEEE J. Emerg. Sel. Topics Power Electron.,
vol. 7, no. 4, pp. 2130–2146, Dec. 2019.
[8] W. Zhang, A. Tarraso, J. Rocabert, A. Luna, J. I. Candela, and P. Rodriguez,
‘‘Frequency support properties of the synchronous power control for
grid-connected converters,’’ IEEE Trans. Ind. Appl., vol. 55, no. 5,
pp. 5178–5189, Sep. 2019.
[9] J. Rocabert, A. Luna, F. Blaabjerg, and P. Rodríguez, ‘‘Control of power
converters in AC microgrids,’’ IEEE Trans. Power Electron., vol. 27,
no. 11, pp. 4734–4749, Nov. 2012.
[10] S. Yazdani, M. Ferdowsi, M. Davari, and P. Shamsi, ‘‘Advanced current-
limiting and power-sharing control in a PV-based grid-forming inverter
under unbalanced grid conditions,’’ IEEE J. Emerg. Sel. Topics Power
Electron., vol. 8, no. 2, pp. 1084–1096, Jun. 2020.
[11] A. Tayyebi, D. Groß, A. Anta, F. Kupzog, and F. Dörfler, ‘‘Frequency
stability of synchronous machines and grid-forming power converters,’’
IEEE J. Emerg. Sel. Topics Power Electron., vol. 8, no. 2, pp. 1004–1018,
Jun. 2020.
93378
[12] M. G. Taul, X. Wang, P. Davari, and F. Blaabjerg, ‘‘Current limiting
control with enhanced dynamics of grid-forming converters during fault
conditions,’’ IEEE J. Emerg. Sel. Topics Power Electron., vol. 8, no. 2,
pp. 1062–1073, Jun. 2020.
[13] P. Rodriguez, C. Citro, J. I. Candela, J. Rocabert, and A. Luna, ‘‘Flexible
grid connection and islanding of SPC-based PV power converters,’’ IEEE
Trans. Ind. Appl., vol. 54, no. 3, pp. 2690–2702, May 2018.
[14] A. Tarraso, N.-B. Lai, G. N. Baltas, and P. Rodriguez, ‘‘Power quality
services provided by virtually synchronous FACTS,’’ Energies, vol. 12,
no. 17, p. 3292, 2019. [Online]. Available: https://www.mdpi.com/1996-
1073/12/17/3292
[15] P. R. Cortes, J. I. C. Garcia, J. R. Delgador, and R. Teodorescu, ‘‘Virtual
controller of electromechanical characteristics for static power convert-
ers,’’ Europen Patent 2 683 075 A1, Jan. 8, 2014.
[16] M. Pérez-Ortiz, S. Jiménez-Fernández, P. Gutiérrez, E. Alexandre,
C. Hervás-Martínez, and S. Salcedo-Sanz, ‘‘A review of classification
problems and algorithms in renewable energy applications,’’ Energies,
vol. 9, no. 8, p. 607, 2016. [Online]. Available: https://www.mdpi.com/
1996-1073/9/8/607
[17] G. N. Baltas, C. Perales-González, P. Mazidi, F. Fernandez, and
P. Rodriguez, ‘‘A novel ensemble approach for solving the transient sta-
bility classification problem,’’ in Proc. 7th Int. Conf. Renew. Energy Res.
Appl. (ICRERA), Oct. 2018, pp. 1282–1286.
[18] J. J. Q. Yu, D. J. Hill, A. Y. S. Lam, J. Gu, and V. O. K. Li, ‘‘Intelligent
time-adaptive transient stability assessment system,’’ IEEE Trans. Power
Syst., vol. 33, no. 1, pp. 1049–1058, Jan. 2018.
[19] P. McNabb, D. Wilson, and J. Bialek, ‘‘Classification of mode damping and
amplitude in power systems using synchrophasor measurements and clas-
sification trees,’’ IEEE Trans. Power Syst., vol. 28, no. 2, pp. 1988–1996,
May 2013.
[20] F. Sulla, E. Måsbäck, and O. Samuelsson, ‘‘Linking damping of elec-
tromechanical oscillations to system operating conditions using neural
networks,’’ in Proc. IEEE PES Innov. Smart Grid Technol., Eur., Oct. 2014,
pp. 1–6.
[21] P. Kundur, N. J. Balu, and M. G. Lauby, Power System Stability and
Control (The Epri Power System Engineering). New York, NY, USA:
McGraw-Hill, 1994.
[22] R. Badar and J. Shair, ‘‘Adaptive NeuroFuzzy sliding mode based damping
control for SSSC,’’ in Proc. SAI Intell. Syst. Conf. (IntelliSys), Y. Bi,
S. Kapoor, and R. Bhatia, Eds. Cham, Switzerland: Springer, 2018,
pp. 91–106.
[23] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Classification, 2nd ed.
Hoboken, NJ, USA: Wiley, 2001.
[24] S. Theodoridis and K. Koutroumbas, Pattern Recognition, 4th ed.
New York, NY, USA: Academic, 2008.
[25] L. Breiman, ‘‘Random forests,’’ Mach. Learn., vol. 45, no. 1, pp. 5–32,
Oct. 2001, doi: 10.1023/A:1010933404324.
[26] S. Shalev-Shwartz and S. Ben-David, Understanding Machine Learning:
From Theory to Algorithms. Cambridge, U.K.: Cambridge Univ. Press,
2014.
[27] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel,
M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas,
A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay,
‘‘Scikit-learn: Machine learning in Python,’’ J. Mach. Learn. Res., vol. 12,
pp. 2825–2830, Nov. 2011.
[28] K. Diakos, Q. Wu, and A. H. Nielsen, ‘‘Phasor measurement unit and
phasor data concentrator test with real time digital simulator,’’ in Proc.
IEEE PES Asia–Pacific Power Energy Eng. Conf. (APPEEC), Dec. 2014,
pp. 1–5.
[29] T. S. Ustun, S. M. Farooq, and S. M. S. Hussain, ‘‘A novel approach for
mitigation of replay and masquerade attacks in smartgrids using IEC 61850
standard,’’ IEEE Access, vol. 7, pp. 156044–156053, 2019.
[30] T. S. Ustun, S. M. Farooq, and S. M. S. Hussain, ‘‘Implementing secure
routable GOOSE and SV messages based on IEC 61850-90-5,’’ IEEE
Access, vol. 8, pp. 26162–26171, 2020.
[31] M. Abdollahi, J. I. Candela, J. Rocabert, M. A. Elsaharty, and P. Rodriguez,
‘‘Novel analytical method for dynamic design of renewable SSG SPC unit
to mitigate low-frequency electromechanical oscillations,’’ IEEE Trans.
Power Electron., vol. 35, no. 7, pp. 7532–7544, Jul. 2020.
[32] A. Stativă and F. Gonzalez-Longatt, ‘‘Peer-to-peer (P2P) MATLAB–
PowerFactory communication: Optimal placement and setting of power
system stabilizer,’’ in Advanced Smart Grid Functionalities Based on
PowerFactory. Cham, Switzerland: Springer, 2018, pp. 301–318, doi:
10.1007/978-3-319-50532-9_12.
VOLUME 8, 2020
G. N. Baltas et al.: Grid-Forming Power Converters Tuned Through Artificial Intelligence
GREGORY N. BALTAS (Student Member, IEEE)
received the B.S. degree in electrical engineering
from the Technological Educational Institute of
Central Greece, Chalkis, Greece, in 2015, and the
M.S. degree in power engineering from the Univer-
sity of Sydney, Sydney, Australia. He is currently
pursuing the Ph.D. degree in data science with the
Universidad Loyola Andalucía,
Since 2018, he has been a Research Assistant
with the Loyola Institute of Science and Tech-
nology (Loyola.TECH), Universidad Loyola Andalucía. His research inter-
ests include applied artificial intelligence, power system stability, and
autonomous energy systems.
NGOC BAO LAI (Student Member, IEEE)
received the B.S. degree in electrical engineering
from the Danang University of Science and Tech-
nology, Da Nang, Vietnam, in 2014, and the M.S.
degree in electrical and information engineering
from the Seoul National University of Science and
Technology, Seoul, South Korea, in 2017. He is
currently pursuing the Ph.D. degree in electrical
engineering wtih the Universitat Politècnica de
Catalunya (UPC).
Since 2018, he has been a Research Assistant with Universidad Loyola
Andalucía. His research interests include robust control, power electronics,
and distributed generation systems.
LEONARDO MARIN (Student Member, IEEE)
received the B.S. degree in industrial engineer-
ing from Cuenca University, Cuenca, Ecuador,
in 2005, and the M.S. degree in automatic
systems and industrial electronics engineering
from the Polytechnic University of Catalonia
(UPC), Barcelona, Spain, in 2015. He is cur-
rently pursuing the Ph.D. degree in electrical engi-
neering with the Renewable Electrical Energy
System Research Centre, Universitat Politècnica
de Catalunya (UPC). Since 2019, he has been a Research Assistant with
Universidad Loyola Andalucía. His research interests include the dynamic
and stability analyses of power systems with high-penetration of renewable
energies, HVDC and hybrid systems, distributed generation systems, and
microgrids.
VOLUME 8, 2020
ANDRÉS TARRASÓ (Student Member, IEEE)
received the M.Sc. degree in electrical engineer-
ing from the Technical University of Catalonia,
Barcelona, Spain, in 2017. He is currently pursu-
ing the Ph.D. degree with the Technical Univer-
sity of Catalonia, Barcelona. From 2017, he has
been a Research Assistant with the Department
of Electrical Engineering, Technical University of
Catalonia, where he has been a Researcher and an
Assistant Professor with the Department of Elec-
trical Engineering. His current research interests include power electronics,
photovoltaics, wind energy systems, and microgrids.
PEDRO RODRIGUEZ (Fellow, IEEE) received
the M.Sc. and Ph.D. degrees in electrical engi-
neering from the Technical University of Catalonia
(UPC), Spain, from 1994 and 2004, respectively.
He was a Postdoctoral Researcher with CPES, Vir-
ginia Tech, USA, the Department of Energy Tech-
nology, Aalborg University (AAU), Denmark, and
the MIT Energy Initiative (MITie), Boston, USA.
He was a Co-Supervisor of the Vestas Power Pro-
gram, Denmark, from 2007 to 2011. He was a
Director of the technology on modern power systems at Abengoa Research,
from 2011 to 2017. Since 2017, he has been a Full Professor with the Loyola
University Andalucia, where he is the Head of LOYOLA.Tech, leading a
Research Program on Intelligent Energy Systems. He is also linked with
the UPC as a part-time Professor. He is in the Clarivate’s list of Highly
Cited Researchers in Engineering, from 2015 to 2018. He has coauthored
one Wiley-IEEE book, more than 100 articles in ISI technical journals, and
around 300 articles in conference proceedings. He is the holder of 16 licensed
patents. He has participated in more than 50 projects with industrial partners
and several EU projects. He is a IEEE Fellow for his contributions in the
control of distributed generation. He has served as an Associate Editor of the
IEEE TRANSACTION ON POWER ELECTRONICS, the IEEE JOURNAL ON EMERGING
AND SELECTED TOPICS ON POWER ELECTRONICS, the IEEE JOURNAL ON INDUSTRIAL
ELECTRONICS AND ENERGIES. His research interests include intelligent energy
systems, distributed generation, and universal energy access.
93379